Humanizer test runs
Humanizer test runs on this site are listed here by ID and date, and each run is 12 tools on the same 24 AI-written inputs and 6 human controls, read on five named detectors in one week. Runs happen every second month. The first run is scheduled, and its page will appear in this hub when it publishes.
A run page is where the numbers on the rest of the site come from. A review, a comparison or a ranking that states a meaning score, an error rate or a detector reading names the run it draws on and links the run page, which in turn links the raw files. Nothing on the site carries a score that cannot be traced to a run in this hub.
What is a run?
One run is every tool in the test set, on the same input set, in the same week, checked on the same detectors, scored with the same rubric. A run gets an ID in the form R1-2026-12, a page of its own under this hub, and a folder of raw files. The full protocol is on how we test AI humanizers; the short version is below.
| Part of a run | What it is |
|---|---|
| Input set | 24 AI-written texts across six genres, four per genre, plus 6 human-written controls, one per genre, fixed for a calendar year |
| Tools | The 12 tools on the main ranking, each on its default mode plus its strongest documented mode where one exists |
| Plan used | The cheapest plan that accepts the input length, recorded with price and checked-on date |
| Detectors | GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, each read on the run date with a screenshot |
| Scoring | Meaning kept 35, errors introduced 20, detector pass rate 30, limits and access 15, to a headline score out of 100 |
| Raters | Two editors blind to the tool rate meaning and count errors; a disagreement of 2 or more points goes to a third read |
What does each run page publish?
Every run page has the same sections in the same order, so a reader who has read one run can read the next without a key.
| Section | What it holds |
|---|---|
| Run ID and date | The ID, the week the tools were run and the week the detectors were read, and the line "Tested: Month Year. Page updated: date" |
| Inputs | The 24 AI inputs and 6 controls by ID and genre, with the generation prompts |
| Tools and versions | Each tool, the mode names from its own interface on the run date, the plan used, its price and the checked-on date |
| Detector readings | Every reading for every input and output, including the baseline row before humanizing, per detector and pooled |
| Scores | Headline score and all four sub-scores for each tool, from the same data file the rest of the site reads |
| Run notes | Tool changes since the last run, errors and retries, and any detector version that could be identified |
| Raw files | inputs.csv, outputs.csv, detector-readings.csv, ratings.csv from both raters, and a screenshot archive, served from this domain |
The raw files are the product, published under the data licence in the terms of use. The tables on a run page are a reading of them, and anyone can check a rating against the output text in outputs.csv or a detector reading against its screenshot.
How often do humanizer test runs happen?
Every second month. The cadence is fixed so that detector model changes, which happen without notice, are caught within a known window, and so that a tool that changes its modes or plans is re-read on a known date. The input set stays fixed for a calendar year; a new set starts a new series, and scores are compared only within a series.
A run is never edited after publication. If a vendor disputes a reading or an editor finds an error, the dispute and the response are recorded in the corrections log and as a dated note on the run page. If a number changes, the change is a new run ID, never a silent edit to the old one.
When does the first run publish?
The first run, R1-2026-12, is scheduled for December 2026. Its page will appear under /runs/ when it publishes, and the revision log on the method page will carry the same entry on the same day. Until then, every tool page on the site states its facts from vendor pages and published studies, and names no score.
| Run ID | Tested | Page |
|---|---|---|
| R1-2026-12 | Scheduled, December 2026 | Appears here on publication |
Every page here will be scored on the same published weights: meaning kept 35, errors introduced 20, detector pass rate 30, limits and access 15. Price is reported, never scored. How these pages are scored
